Received January 12, 2021, accepted January 26, 2021, date of publication February 9, 2021, date of current version February 19, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3058066

CoAID-DEEP: An Optimized Intelligent

Framework for Automated Detecting

COVID-19 Misleading Information

on Twitter

DIAA SALAMA ABDELMINAAM

1,2

, FATMA HELMY ISMAIL

, MOHAMED TAHA

AHMED TAHA

, ESSAM H. HOUSSEIN

, AND AYMAN NABIL

Faculty of Computers and Artiﬁcial Intelligence, Benha University, Benha 13511, Egypt

Faculty of Computer Science, Misr International University, Cairo 11341, Egypt

Faculty of Computers and Information, Minia University, Minia 61519, Egypt

Corresponding author: Diaa Salama Abdelminaam (diaa.salama@fci.bu.edu.eg)

This work was supported in part by Faculty of Computers Science, Misr International University - Grant number 28211231302952.

ABSTRACT COVID-19 has affected all peoples’ lives. Though COVID-19 is on the rising, the existence of

misinformation about the virus also grows in parallel. Additionally, the spread of misinformation has created

confusion among people, caused disturbances in society, and even led to deaths. Social media is central

to our daily lives. The Internet has become a signiﬁcant source of knowledge. Owing to the widespread

damage caused by fake news, it is important to build computerized systems to detect fake news. The paper

proposes an updated deep neural network for identiﬁcation of false news. The deep learning techniques

are The Modiﬁed-LSTM (one to three layers) and The Modiﬁed GRU (one to three layers). In particular,

we carry out investigations of a large dataset of tweets passing on data with respect to COVID-19. In our

study, we separate the dubious claims into two categories: true and false. We compare the performance of the

various algorithms in terms of prediction accuracy. The six machine learning techniques are decision trees,

logistic regression, k nearest neighbors, random forests, support vector machines, and naïve Bayes (NB).

The parameters of deep learning techniques are optimized using Keras-tuner. Four Benchmark datasets were

used. Two feature extraction methods were used (TF-ID with N-gram) to extract essential features from the

four benchmark datasets for the baseline machine learning model and word embedding feature extraction

method for the proposed deep neural network methods. The results obtained with the proposed framework

reveal high accuracy in detecting Fake and non-Fake tweets containing COVID-19 information. These results

demonstrate signiﬁcant improvement as compared to the existing state of art results of baseline machine

learning models. In our approach, we classify the data into two categories: fake or nonfake. We compare

the execution of the proposed approaches with Six machine learning procedures. The six machine learning

procedures are Decision Tree (DT), Logistic Regression (LR), K Nearest Neighbor (KNN), Random Forest

(RF), Support Vector Machine (SVM), and Naive Bayes (NB). The parameters of deep learning techniques

are optimized using Keras-tuner. Four Benchmark datasets were used. Two feature extraction methods were

used (TF-ID with N-gram) to extract essential features from the four benchmark datasets for the baseline

machine learning model and word embedding feature extraction method for the proposed deep neural

network methods. The results obtained with the proposed framework reveal high accuracy in detecting Fake

and non-Fake tweets containing COVID-19 information. These results demonstrate signiﬁcant improvement

as compared to the existing state of art results of baseline machine learning models.

INDEX TERMS Fake news, COVID-19, misleading information, pandemic, social media, deep learning.

The associate editor coordinating the review of this manuscript and

approving it for publication was Pasquale De Meo.

I. INTRODUCTION

COVID-19 is rumored to be caused by a new SARS-CoV,

which ﬁrst appeared in China in December 2019 and

27840

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/

VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

soon spread. The Zika virus outbreak was declared a public

health emergency of international signiﬁcance on January 30,

2020, and the virus was named COVID-19 in March of the

same year [1]. According to WHO, as of May 6, 2020, more

than 3.5 million cases of COVID-19 have been reported to the

World Health Organization. The most common symptoms of

CVID-19 infection include cough, trouble breathing, fever,

sore throat, and an inability to taste or smell [2].

While in the midst of the COVID-19 scenario, complexities

relating to COVID-19 have risen and triggered signiﬁcant

social disturbances. On the other hand, fraudulent purveyors

of the COVID-19 commodity have caused many people to

suffer. For example, an Arizona man was dead and his wife

was hospitalized after the couple ingested a form of Chloro-

quine to prevent COVID-19. On the other hand, poor contact

is adversely affecting social order [1].

Misinformation is one type of the many forms of misinfor-

mation such as rumors, misleading content on the internet,

and fake news. Many studies have described fake news as

‘‘news articles that are intentionally written to mislead or

misinform readers, but can be veriﬁed as false by means of

other sources’’ Fake news is due to its exposure to public

polarization and quarrelling [3], [4].

Various examples are seen in both the 2016 US [5] pres-

idential campaign and the 2019 Indian airstrike in Balakot.

Accurately distinguishing between genuine news and fake

news will be needed for building intelligent AIs. Social net-

working stages, including Facebook, Twitter, etc., are strug-

gling to cope with the amount of misinformation that is being

shared on these sites.

Fake news can usually be divided into three distinct cate-

gories false stories, large scale hoaxes, and satirical fabrica-

tions [6]. The aim of the fabricated interviews and malicious

intent category is to expose fakes that thrive on social media.

Large scale hoaxes seem as if they are reality but are just a lie.

The ﬁnal category are satirical news intended to amuse users

and are frequently disguised as real news by authors. The

objectivity of news sites is the most critical aspect in assess-

ing their reliability. Malicious web sites adopt the domains

of popular, trustworthy websites. Fake news websites can

be used to survey readers and create a data collection for

research. Unfortunately, false news may easily be identiﬁed

on trustworthy websites by error. To gather real and false

news stories, both honest and fake articles must be collected.

Humans are important to detect the accuracy of the news.

Alternative veriﬁcation approaches are not the only means

of validating news. crowdsourcing and computational fact

checking models are used to annotate dubious news and to

provide fact-checking information [7].

A benchmark dataset of fake news has not been agreed

until now because of the trouble of collecting fake news and

the ambiguity in providing a clear deﬁnition [7]. However,

some authors created a dataset from the statements collected

from social media such as LIAR [8]. Some authors altered

Wikipedia sentences to produce statements and provide evi-

dence for or against such claims in Wikipedia articles [9].

Another fake news dataset is collected from Facebook and

Twitter [10]–[13]. Finally, the complete dataset is provided

in [14], where the authors provided a dataset that contains

information about the content and the social context of the

news. Researches on fake news make a difference between

content features and context features. The content features

of fake news are linguistic features. The context features of

fake news are the surrounding information such as user’s

characteristics and social network-based features [7].

Therefore, there is a high degree of risk publishing fake

news over social media. The big truth of news is the need of

the hour, and it’s something that we must ﬁx.

Recently, there are many examples for fake news in France

and the USA such as fake news about the Presidential candi-

dates during the France and US presidential election, which

was shared over thousands of times, and spread quickly. The

following examples of fake news that spread quickly and how

content broadcasts over social media

• The ﬁrst example of fake news is titled by the

French presidential campaign commercial, sponsored by

Saudi Arabia. This is a paper published on February 24,

2017 that argues that French presidential candidate

Emmanuel Macron (a centrist candidate) was sponsored

by Saudi Arabia (30% of Macron’s campaign funded

by Saudi Arabia during France presidential election.).

The characteristic indicated that The story is fabricated.

A fake site was created to mimic the real site of Le Soir,

and to spread false information. The story on Facebook

has received over 10,000 likes, shares, and comments.

• The second example of Fake News, Titled by Hillary

Clinton, deletes Hillary Clinton from Twitter. This

fake news adversely affects how the US presidential

elections turn out. punctuation like commas, apostro-

phes, quotations, question marks and more are omitted,

to decrease the model’s computational cost and increase

its efﬁciency.

• The third example of fake news (a tweet circulated

over 1,700 times), titled by Marine Le Pen, mocked

the ‘‘Masha and the Bear’’ cartoon because the little

girl in the story wears a veil. The paper was released

on February 26, 2017 and reported that French presi-

dential candidate and National Front chief Marine Le

Pen mocked a children"s cartoon, Masha and the Bear,

because Masha wore a ‘‘veil.’’ The truth revealed that

Marine Le Pen did not tweet that - a short video of

the image is doctored. Secretnews.fr published an article

discussing the subject in 2014.

• One of four False News stories, The French state,

will spend 100 million euros purchasing hotels on

housing migrants. It is announced in the news on

March 10th 2017 that the Council of Europe Develop-

ment Bank (CEB) will lend the French state 100 mil-

lion euros to buy hotels to house asylum seekers. The

truth revealed that the claim was misleading and two

independent news stories had been conﬂated and altered

to exaggerate the assistance that was given to asylum

VOLUME 9, 2021 27841

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

seekers. The connection has been shared more than

10,000 times on Facebook. stemming is replacing the

sufﬁx with a root word to reduce the number of word

types in the results. For example, ‘‘Making,’’ will be

represented by the word ‘‘Made,’’ and ‘‘Maker’’ will be

represented by the word ‘‘make.’’

• Finally, consider a real tweet by US President Trump.

He has ordered a funding freeze on the World Health

Organization, accusing it of practicing ‘‘blackout’’ over

the spread of the Coronavirus. He found fault with

the WHO for not being able to tackle the emerging

COVID-19 crisis, which caused the deaths of thou-

sands. The World Health Organization had given mis-

leading information gathered from China, which had

spread the virus across its territory. ‘‘He also considered

that the international organization’’ Most publications

declined to offer the same standard of transparency as

this one.

More and more, false allegations have been made on social

media that harm the reputations of politicians. Some recent

examples have suggested that lawmakers and elected ofﬁcials

indulge in bad conduct but the truth is that the claims are

unfounded. Some outlets describe politicians as heroes for

tasks they did not achieve. Either way, fake news adversely

affects public conﬁdence in the media. Furthermore, tall tales

can affect people’s opinions.

The previous examples contribute to the issue of people not

knowing because the consumer is misinformed. The negative

inﬂuence of social media misinformation has a widespread

negative impact on society. When people spread false facts,

they negatively affect people’s emotions. We made an attempt

to model the problem in our proposed model.

The paper contributions of the proposed techniques can be

summarized as follows:

1) We have the ﬁrst initiative to apply deep learning

techniques to the COVID-19 dataset to detect false

news.

2) We proposed a novel Fake News Detection system

on social media platforms for COVID-19 dataset and

others using Modiﬁed Deep Neural Network methods.

3) We conduct a systematic experiment using various

state-of-the-art machine learning algorithms to deter-

mine the efﬁciency of the proposed deep learning

algorithms.

4) In Fake News Detection, the proposed algorithm

achieved 98.57% accuracy on the best dataset.

The remainder of this paper is organized as follows: the

related work is presented in section 2. The proposed method-

ology is introduced in section 3. The experiment results and

discussion are discussed in section 4 and section 5, respec-

tively. Finally, conclusions are presented in section 6.

II. RELATED WORK

This section introduces different machine learning techniques

to detect fake news in general. The ﬁrst subsection is related

to COVID-19 fake news detection since it is the most recent

topic.

A. COVID-19 FAKE NEWS DETECTION

Since the appearance of the ﬁrst COVID-19 case on

December 31, the World Health Organization (WHO)

declared it as a pandemic emergency. Social media news

and tweets contain information or misinformation about

COVID-19. Ordinary people become more anxious to read

more to know how to protect themselves. The authors in [25]

analyzed the sources of COVID-19 misinformation. Their

analysis revealed that most of the misinformation about

COVID-19 are fabricated from true information rather than

invented. Detecting fake news about COVID-19 attracted data

scientists. The authors in [22] applied 10 machine learning

algorithms, with 7 feature extraction techniques to detect

whether the corpus of news is fake or real. They tested their

proposed classiﬁer on 3,047,255 COVID-19 related tweets.

The best performance measures are achieved by NN, DT, and

LR classiﬁers. In [26], the authors extracted the textual fea-

tures of COVID-19 tweets beside user and network features.

They proposed mBERT (multilingual Bidirectional Encoder

Representations from Transformers) which is a deep neural

network approach. mBERT achieved the highest performance

measures compared with traditional machine learning tech-

niques such as SVM, RF and a multilayer perceptron. In [20],

the authors applied two pipelined pre-trained deep learning

natural language frameworks named BERT and ALBERT.

They used a public dataset on COVID-19 that contains more

than 5000 COVID-19 false claims. Their proposed model

yielded the best performance results.

NLP researchers have been working on developing algo-

rithms for the detection of online COVID-19 related disin-

formation. To develop any algorithm, we require a corpus.

So members of the NLP community created the various fake

news datasets: FakeCovid [15], ReCOVery [16], CoAID [1],

and CMU-MisCOV19 [17]. Yichuan Li et al. [18] developed

multi-dimensional and multilingual MM-COVID corpora,

which covers six languages. Mabrook et al. [19] created a

large Twitter dataset related to COVID-19 misinformation.

And authors developed an ensemble-stacking model with

six machine learning algorithms on the created dataset for

detecting misinformation.

Elhadad et al. [22] constructed a voting ensemble machine

learning classiﬁer for fake news detection that uses seven

feature extraction techniques and ten machine learning mod-

els. Tamanna et al. [21] used the COVIDLIES dataset

to detect the misinformation by retrieving the misconcep-

tions relevant to the Twitter posts. For COVID-19 fake

news detection and fact-checking, Rutvik et al. [20] pro-

posed a two-stage transformer model. The ﬁrst model

retrieves the most relevant facts about COVID-19 by using

a novel fact-checking algorithm, and the second model,

by computing the textual entailment, veriﬁes the level of

truth.

27842 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 1. Comparison between different Fake News Detection for COVID-19.

also Fake news can be found in form of ‘‘fake cures’’

such as in [23], [24] that point out it inﬂuences the decision-

making process in medicine

Adapting all these classical and hybrid related work tech-

niques, we developed a COVID-19 fake news detection

system in this paper.

B. FAKE NEWS DETECTION ALGORITHMS

Detecting fake news becomes one of the most critical

tasks of artiﬁcial intelligence scientists. There are two main

approaches for detection: the machine learning approach and

the Deep Learning approach.

1) Machine learning approach for fake news

detection: Deception has been studied and deﬁned

as the creation of a false conclusion by transmitting

a false message. In their study, [27] analyzed a set

of linguistic features and investigated three classi-

ﬁers. SVM achieved the highest precision, recall, and

F-measure. However, linguistic features and visual fea-

tures are commonly used in SVM approaches such as

[27]–[31]. In [32], authors distinguished between fake

news, satire news, and real ones by introducing a set

of distinguishing features such as the titles. In [33],

both content and context-based features were used to

detect fake news using a Decision Tree(DT). How-

ever, Random Forest (RF) and Decision Tree (DT)

were applied using user characteristics in [34], [35] to

detect the trustworthiness of users writing the news.

Additional features like topic models based features are

used in [33]. In [29], the authors deﬁned some linguistic

cues of deception and applied Random Forest (RF)

to detect fake news. Logistic Regression (LR) has

shown competitive performance in detecting fake news

in [12], [29], [36].

2) Deep learning approach for fake news detection:

Deep learning classiﬁers have become popular in

recent years. The approach of deep learning is efﬁ-

cient in terms of extracting relevant features [37].

Recurrent neural networks (RRNs) and, in particular,

LSTM is efﬁcient in modeling sequential data [12].

In [38], the authors proposed different RNN architec-

tures, namely tanh-RNN, LSTM, and Gated Recurrent

Unit (The Modiﬁed GRU), and The Modiﬁed GRU

achieved the best performance. In [39], LSTM has been

fed by a mix of content and context-based features of

news, and it achieved good accuracy in detecting fake

news. CNN’s are a class of neural networks that gain

popularity in the NLP ﬁeld [40]. In [41], both RNN

and CNN are used to detect false news and show a

better performance than the performance of baselines.

In [42], the authors used LSTM and hybrid LSTM-

CNN architectures. The simplest LSTM showed the

best performance. In [8], a hybrid model of RNNs and

CNNs was used by the authors where the text informa-

tion is encoded via CNN, and LSTM encodes the meta-

data of the author. This hybrid model outperformed

VOLUME 9, 2021 27843

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 2. Comparison between different Fake News frameworks.

the baseline model. Table 2 represents a summary of

related work

III. METHODOLOGY

The proposed system of fake news detection consists of two

main categories, as shown in Figure

1. The ﬁrst category uses

regular machine learning algorithms, and the second category

by using deep neural networks. The ﬁrst category detects fake

news using six baseline traditional machine learning tech-

niques. The machine learning techniques are decision tree

(DT), logistic regression (LR), K-nearest neighbor (KNN),

random forest (RF), support vector machine (SVM), and

Naive Bayes (NB). The second category detects fake news

using two proposed deep learning techniques. The two deep

learning techniques are Modiﬁed LSTM (one to three lay-

ers) and Modiﬁed GRU (one to three layers). In the ﬁrst

step (preprocessing step) for the two categories; includes

removing unimportant characters, tokenization, removing

stop wording, and stemming. In the third step, the feature is

extracted using TF-IDF with N-grams for the ﬁrst category,

ML techniques. In contrast, for the second category, i.e., deep

learning techniques, the feature is extracted using the word

embedding method with Glove to build a word embedding

matrix. In the fourth step, the parameters of traditional

machine learning techniques are optimized using a grid

search with stratiﬁed cross-validation, while the parameters

of deep learning techniques are optimized using a Keras-tuner

library. The performance of each technique is evaluated by

measuring accuracy, precision, recall, and F-Measure. Each

step is described in detail in the following subsections.

A. DATA COLLECTION

Experiments were conducted using Four Twitter fake news

datasets in different topics. the ﬁrst topic is CoAID

(COVID-19 heAlthcare mIsinformation Dataset) [1]. the

secand dataset include disasters [45], PolitiFact [46], gossip

27844 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

FIGURE 1. The proposed Framework for Automated Detecting COVID-19 Misleading

information onTwitter.

cop [46]. The disaster dataset is collected from Kaggle about

the topic disaster, while the second and third datasets are

related PolitiFact and gossip cop topics collected from Fake-

NewsNet.

1) CoAID (COVID-19 heAlthcare mIsinformation

Dataset) is a diverse array of COVID-19 healthcare

disinformation, from fake news on blogs and social

media, along with the impact that individuals have on

VOLUME 9, 2021 27845

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 3. The disaster dataset description.

TABLE 4. The final PolitiFact and final gossip cop datasets description.

such fake news. The data contains 4,251 news men-

tions, 296,000 user engagements, 926 tweets referenc-

ing the COVID-19 and ground truth label.

2) The disaster dataset [45] has ﬁve features; id, text,

location, keyword, and target (see Table [

3]). The Dis-

aster dataset has a text of 7613 tweets. Each given

tweet is about a real disaster or not labeled as 1 and 0,

respectively. In particular, 4342 tweets show a real dis-

aster, while 3271 shows not. In our experiment, we have

used two features, which are text and target, as a label,

which shows whether a tweet is about a real disaster (1)

or not (0).

3) The PolitiFact dataset PolitiFact [46] has two ﬁles

1) politifact_real.csv, which contains samples related

to real news that includes 432 tweets, 2) politi-

fact_fake.csv contains samples related to fake news that

includes 618 tweets. We merged politifact_real.csv and

politifact_fake.csv ﬁles into one ﬁle where each tweet

belongs to politifact_real labeled as 0 while each tweet

belongs to politifact_fake labeled as 1. The ﬁnal Poli-

tiFact dataset has ﬁve features: id, URL, title, tweet-id,

and label (see Table ([

4]). In our experiment, we used

the title to represent the text of the tweet and label

features.

4) The gossip cop dataset [46] has two ﬁles 1) gos-

sip cop_real.csv, which contains sample tweets related

to real news that includes 5328 tweets, 2) gos-

sipcop_fake.csv contains sample tweets related to

fake news. We selected 5322 tweets from gossip-

cop_fake.csv. We merged gossipcop_real.csv and gos-

sipcop_fake.csv ﬁles into one ﬁle where each tweet

belongs to gossipcop_real labeled as 0 while each tweet

belongs to gossipcop_fake labeled as 1. The ﬁnal gos-

sip cop dataset has ﬁve features, including id, URL,

title, tweet-id, and label (see Table [

4]). In our experi-

ment, we used the title to represent the text of the tweet

and label features.

B. DATA PREPROCESSING

Data preprocessing is an important phase for any senti-

ment analysis system, especially for social media content.

Twitter data is the popular unstructured datasets collected

of information from people entered his/her feelings, opin-

ion, attitudes, products review, emotions, etc. These datasets

need to be subjected to certain reﬁnements by performing

preprocessing techniques to the next phases of elaboration,

i.e., applying ML/DL techniques. The basic cleaning opera-

tions within preprocessing techniques used in this work are

removing unimportant characters, stop-word removal, tok-

enization, a lower casing, sentence segmentation, and punc-

tuation removal. They will help us to reduce the size of actual

data by removing the irrelevant information that exists in the

data and then to achieve better performances. In our study,

the preprocessing involves a series of techniques which are

listed in the following steps:

• Lower casing: simply, it one of the basic cleaning

operations to convert a word to lower cases such as

NLP -> nlp.

• Removing unimportant data: the punctuation like

commas, apostrophes, quotes, question marks, and more

which do not add much value to a natural language

model are deleted.

• Tokenization: It is the key aspect of working with

text data to separate a piece of text into smaller units

called tokens. The tokens are including paragraphs and

sentences which can be further broken into words. For

example, consider this sentence before tokenization:

‘‘never give up’’, after tokenization it comes ‘never’,

‘give,’ ‘up’.

• Removal of Stop Word: a stop word usually refers to

the most common words in a language that does not

add much meaning to a sentence such as articles, prepo-

sitions and conjunctions, and some pronouns. These

words are removed from each tweet with the datasets.

• Stemming: stemming is removing the sufﬁx from and

transform it to its root word to reduce the number of

word types or classes in the data. For example, the words

‘‘Making,’’ ‘‘Made,’’ and ‘‘Maker’’ will be reduced to

the word ‘‘make.’’

C. DATA SPLITTING

In this step, a dataset from 90 percent of the training set and

another dataset from 10 percent of the testing set are used.

The training set is fed into the ML/DL models to ﬁnd out

what should be done with the data, and the unseen test set is

used as a check on the results.

D. APPLYING OPTIMIZATION AND LEARNING MODELS

In this step, we applied using two Learning models (Regular

Machine Learning algorithms, Deep Learning models). Fur-

ther details about each model are presented as follows:

27846 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

1) REGULAR MACHINE LEARNING

We applied six machine learning models after two steps.

ﬁrstly is by applying Machine Learning Feature Extraction

Method then optimized the models using Hyperparameters

Optimization Methods. Further details about each model are

presented as follows:

1) Machine Learning Feature Extraction Method:

In this step, we have used the TF-IDF feature extraction

method with different sizes of N-Gram and 3000 matrix

size. N-gram is the simplest model that assigns proba-

bilities to sentences and sequences of words beginning

with length n. The value of n can be one (unigram),

two (bi-gram), three (tri-gram) and so on. Various uni-

grams and bigrams can be classiﬁed into two groups:

character-based, and word-based. A speciﬁc set of text

characters taken from a phrase. We are using this tech-

nique since related terms would have a high proportion

of N-grams in common. Typical values for n are 2 or 3;

these correspond to bigrams or trigrams, respectively.

For example, the word computer results in:

• the generation of the bigrams *C, CO, OM, MP,

PU, UT, TE, ER, R*

• o and the generation of the trigrams **C,

* CO, COM, OMP, MPU, PUT, UTE, TER,

ER*, R** Where ‘*’ denotes a padding space.

Character-based N-grams are generally used in

measuring the similarity of character strings. The

Term Frequency-Inverted Document Frequency

(TF-IDF) is a well-known feature method to eval-

uate the importance of a word in a document.

According to this work, the TF-IDF method is

used to measure the importance of a term within

a tweet in the fake news datasets. The key idea of

the TF-IDF method is converting the tweets into

a Vector Space Model (VSM) and then calculating

the importance of the term by counting its frequen-

cies within the tweets. The word-based frequency

is counted using different n-grams, including uni-

gram, bi-gram, and tri-gram, etc.

2) Hyperparameters Optimization Methods: In this

step, we have used the hyperparameters optimization

techniques to select the best value for each param-

eter of regular machine learning models, including

Grid Search with stratiﬁed 10-fold cross-validation

described as follows:

• Grid search is a hyperparameter optimization

technique for hyperparameter tuning, which is

used to methodically select the best value that

achieves the best performances for an ML model.

It evaluates ML model for each combination

of algorithm parameters speciﬁed in a grid

and then reports the optimal values of model

hyperparameters.

• K-Fold Cross-Validation is mainly used for

hyperparameter tuning by dividing the sample of

datasets into a training set to train the model,

and a test set to evaluate it. The dataset is split

into k equal partitions where k-1 groups are used

for training, and the one fold is held for the

testing model. This process is repeated k times

(i.e., k = 10), including one fold, is used for testing

and k-1 folds for the training set. In our experiment,

we used k = 10. In the 10-fold CV process, 90%

of data were used for the training, and 10% of data

were used for testing purposes.

3) Machine Learning Models: We have used six regu-

lar machine learning algorithms, which are Decision

Tree, Random forest, K-Nearest Neighbor, Logistic

Regression, Support Vector Machine, and Naive Bayes,

to classify news into fake and real news. Further details

about each model are presented as follows:

• Decision Tree (DT) [47] are useful supervised

Machine learning algorithms that can perform

our classiﬁcation tasks in this paper. It consists

of nodes and branches, where the tests on each

attribute are represented at the nodes, the outcome

of this procedure is represented at the branches,

and the class labels are represented at the leaf

nodes. The goal is to create a model that classifying

the value of a target variable by learning simple

decision rules concluded from the data features.

• Random forest (RF) [48], [49] is a supervised

machine learning algorithm that uses a collection

of decision trees, providing more ﬂexibility, accu-

racy, and ease of access. This algorithm dominates

over decision trees algorithm as decision trees pro-

vide low accuracy compared to the random for-

est algorithm. In simple words, the random forest

approach increases the performance of decision

trees. It is one of the best algorithms in classiﬁ-

cation techniques, and we used it in our paper. The

goal is to create a model that classifying the value

of a target variable by learning simple decision

rules concluded from the data features from more

than the tree.

• K-Nearest Neighbor (KNN) [50], [51] is a Super-

vised classiﬁcation algorithm. It is one of the most

straightforward and widely used algorithms which

depends on its k value; K speciﬁes the number of

neighbors, and its algorithm is as follows:

– Choose the number K of neighbor.

– Take the K Nearest Neighbor of a new data

point, according to Euclidean Distance. (We can

increase or decrease it as you like to get the best

accuracy that we needed)

– Among the K-neighbors, Count the number of

data points in each category.

– Assign the new data point to a category, where

you counted the most neighbors

• Support vector machine (SVM) [52], [53]

SVM is a supervised learning system used for

VOLUME 9, 2021 27847

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

classiﬁcation problems with an associated algo-

rithm. In this process, each data object is plotted in

n-dimensional space with the value along with the

coordinate deciding the value of the item. Classiﬁ-

cation is achieved by discovering the hyper-plane

best dividing the categories. SVMs are capable of

performing a non-linear classiﬁcation by directly

translating inputs into high-dimensional feature

spaces.

• Logistic Regression (LR) [54] is a Machine

Learning algorithm used for classiﬁcation prob-

lems. It is a predictive analysis algorithm and based

on the concept of probability. It is based on the

sigmoid function where output is the probability

(Value of output ranges from 0 to 1), and input

can be from-inﬁnity to +inﬁnity. If we need to

classify our data into two classes, then if the output

probability range is less than 0.5, then our data in

the ﬁrst class (class tag (0)), and if the probability

range more than 0.5, then our data in the second

class (class tag (1)).

• Naive Bayes (NB) [55] is a probabilistic machine

learning model based on Bayes’ theorem. It makes

classiﬁcations using the Maximum Posterior deci-

sion rules in a Bayesian setting.

2) DEEP LEARNING

We applied Two Deep learning models after two steps. ﬁrstly

is by applying Deep Learning Feature Extraction Method

then optimized the models using Hyperparameters Optimiza-

tion Methods. Further details about each model are presented

as follows:

1) Deep Learning Feature Extraction Method In this

step, we have used word embeddings, which generally

converts text data, i.e., words into vectors. It represents

every word in an n-dimensional dense vector where

similar words will have a similar vector. The more efﬁ-

cient word embeddings techniques which have proven

there capability to convert words into vectors are GloVe

and Word2Vec. According to this work, GloVe [56]

represents the tweets within the fake news datasets into

dense vectors, which fed into the deep learning models.

The gloVe is an unsupervised learning algorithm for

word embeddings, which is used to obtain the vector

representations for words. The key idea of the GloVe

technique is to discover the closeness of two words,

with their separation in a vector space to create vector

representations called word embedding vectors. The

embedding vectors are created by aggregating global

word-word co-occurrence statistics from the datasets

and then resulting in the matrix representations, includ-

ing measuring the closeness of two words in a tweet.

We used glove.twitter.27B.zip that includes a different

dimension of vectors, which are 25d, 50d, 100d, and

200d vectors. We used 200d vectors to build the embed-

ding matrix.

TABLE 5. Hyperparameters configurations selected by Keras-tuner.

FIGURE 2. Deep neural network architecture.

2) Hyperparameters Optimization Method: For hyper-

parameters optimization, we have used a

Keras-tuner [57] library to pick the optimal set of

hyperparameters in hidden layers (Modiﬁed LSTM or

Modiﬁed GRU) and dropout layers. We set different

values for different parameters: the number of neurons,

reg_rate for l2 regularization technique [58], and the

dropout rate for the dropout layers [59]. For this,

we have applied the Keras-tuner on the training dataset

to select the best parameters, as shown in

3) Deep neural network:

2 shows the deep neural net-

work architecture that is used to classify news into

fake and real. It consists of a word embedding matrix

as input to embedding layer, embedding layer, hidden

layers including Long Short-Term Memory (LSTM),

or Gated Recurrent Units (GRU), ﬂatten layer [60], and

output layer. In the word embedding matrix, the GloVe

word embedding technique has been used to calcu-

late word embeddings using a co-occurrence matrix in

between words within fake news tweets, which is called

the embedding matrix. The embedding matrix is used to

represent the tweets into dense vectors. The embedding

layer, hidden layers, and output layer are described as

follows:

• Embedding layers The embedding layer is imple-

mented in the Keras library [61]. Regarding

this work, Keras library is used to initialize the

27848 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

embedding layer to learn an embedding for all

of the words in the training dataset. The Keras

embedding layer has three arguments, including

a) input_dim deﬁnes the size of the vocabulary

in the dataset.

b) output_dim deﬁnes the size of the vector space

in which words will be embedded.

c) input_length deﬁnes the length of input

sequences as deﬁned for any input layer

of a Keras model. The embedding layer

is conﬁgured as follows; input_dim equals

20000 because the number of words is 20000,

output_dim equals 200 and because we used

200d vectors of golvetweet and input_length

equals 32.

• Hidden Layer Two different neural network mod-

els are used; LSTM or GRU model. For each

model, a different number of hidden layers has

been applied, including one, two, and three lay-

ers. Also, one dropout layer and different numbers

of neurons in each hidden layer have been used.

The ReLU (Rectiﬁed Linear Unit) [62] activation

function has been applied for the hidden layers.

For each hidden layer, l2 regularization techniques

have been used by adopting reg_rate value for l2.

Also, we used the dropout layer and the different

number of dropout rate.

• Output Layer The output layer provides the ﬁnal

output of the model where the neural network

model classiﬁes the inputs tweets into two cate-

gories; real or fake. In particular, the output layer

has one neuron, which detected the news within

an input tweet in terms of fake or real. In this

layer, we used the ADAM optimizer [63] and sig-

moid [64] is the activation function.

4) Recurrent Neural Network (RNN):

Problems that cannot be boiled down to a set number

of inputs and outputs. Problems in which the device

is needed to store and use background information.

Hard/impossible to choose the exact meaning of a word

There is always new data available that is longer than

everything else. A recurrent neural network is useful

because it allows for the intermediate values (states) to

store information about past inputs for a time that is

not set a priori [65]. The RNN repeats the same task

for each element in the sequence [66]. The RNN uses

the secret state to store the state of each moment, and

the state depends on the previous moment and current

input. The new secret state effectively capitalizes on

the past knowledge. Thus, an RNN can perform this

function via dynamic processes. The architecture of an

RNN is depicted in Figure

3. Given an input sequence

X = [x

, x

· · x

] of length T, an RNN determines

the hidden state ht at the time t of the sequence as

FIGURE 3. RNN architecture.

FIGURE 4. LSTM architecture.

in equation (1):

= tanh

(

t−1

+ W

+ b

)

(1)

Despite the great value of RNNs for learning sequen-

tial patterns, the gradient descent method is difﬁ-

cult to implement because of the well-known gradient

vanishing/explosion problem [67]. To solve these

two separate issues, research is being performed on

Long Short-Term Memory (LSTM), GRU, etc. Con-

sequently, modiﬁed GRU, and modiﬁed LSTM are

chosen as our methodology.

5) Long Short-Term Memory network (LSTM): Long

Short-Term Memory network (LSTM) [40], [68] is

a deep recurrent neural network which is more reli-

able than the traditional recurrent neural network when

used in tasks with long time lags. The main difference

between an RNN and an LSTM is that an RNN has a

single tanh layer, while an LSTM has four interactive

LSTM layers (see Figure

4). The LSTM memory cell is

composed of a memory block and three multiplicative

gating units. For the gate, the sigmoidal nature of the

function σ ranges from 0 to 1 [69], [70].

LSTM has three of these gates to protect and control

the cell state. The three gated are the

• Forget gate layer f

, as shown in Figure (

6 a). For-

get gate layer is used to decide what information

throw away from the cell state). Forget gate layer as

shown in Equation (

2) Output a number is between

0 and 1.

= σ



[

t−1

, x

]

+ b



(2)

VOLUME 9, 2021 27849

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

• Then Add new information as shown in

Figure (

6 b) is to decide what new information

store in the cell state - Input gate layer i

as shown

in Equation (

= σ

(

[

t−1

, x

]

+ b

)

(3)

then Decides which values we will update as

shown in Figure (

6 c) (Tanh layer as shown in

Equation (

) by creating a vector of new can-

didate values

= tanh

(

[

t−1

, x

]

+ b

)

(4)

• Update cell state C

, as shown in Figure (

6 c): by

Forgetting the things we decided to forget earlier

∗ C

t−1

and Adding information we decide to be

added i

∗

as shown in the following formula (

= f

∗ C

t−1

+ i

∗

(5)

• Create Output as shown in Figure (

6) (Output

gate layer o

, Tanh layer)by Decide what we are

going to Output as shown in (6 d)

– Output gate layer o

as shown in Equation (6):

Decides what parts of the cell state we are going

to Output

= σ

(

[

t−1

, x

]

+ b

)

(6)

– Tanh layer h

: as shown in Equation (

7) Push

the values between −1 and +1

= o

∗ tanh

(

)

(7)

• Peephole as shown in Figure (

6 e) to Let the gate

layer look at the cell state (entire/ partial) as shown

in the following Equations (8)

= σ



[

t−1

, h

t−1

, x

]

+ b



= σ

(

[

t−1

, h

t−1

, x

]

+ b

)

= σ

(

[

, h

t−1

, x

]

+ b

)

(8)

• Coupled forgot and input gates as shown in

Figure

6: f Not deciding separately as shown in the

following Equations (

= f

∗ C

t−1

(

1 − f

)

∗

(9)

So we can summarise the LSTM into four steps

– Step 1: Forget gate layer

– Step 2: Input gate layer

– Step 3: Combine step 1 and 2

– Step 4: Output the cell state

Though LSTM uses a certain kind of RNN. LSTM

learning techniques are able to learn long term

dependencies. Although LSTM cannot learn to ﬁll

a wide gap in knowledge, RNNs do not have a gap

problem. LSTM minimizes the number of losses.

LSTM embraces character sequences of varying

lengths, such so that no linguistic features are

FIGURE 5. GRU architecture.

needed to be extracted [5]. This algorithm is also

compact; the update complexity per weight and

time step and storage complexity per weight is on

the order of O (1)

• Concept of the Gated Recurrent Unit (GRU):

The GRU has the same structure as either a basic

RNN or STM, except the GRU updates the hid-

den state [71], [72].. The key difference between

LSTM and GRU is that LSTM combines forget and

input layers into a single ‘‘update gate’’, combines

cell state and hidden state, and is more conve-

nient and common. Instead of explicitly updating

the current hidden state with the previous hidden

state, GRU uses a reset gate and updates the gate,

deciding if the information in the previous hid-

den state is useful, then retains useful information

and removes useless information [73]. Figure (

shows the architecture of GRU.

The way GRU updates h

is as follows:

a) The reset gate r

and update gate z

as shown

in Equation (10) and in Equation Equation (11)

= σ

(

t−1

+ W

+ b

)

(10)

= σ

(

t−1

+ W

+ b

)

(11)

For the gate, σ is a logistic sigmoid, The reset

gate rt, and update gate zt ranging from 0 to 1.

b) Candidate hidden state

= tanh



(

∗ h

t−1

)

+ W

+ b



(12)

The candidate hidden state (

) is shown in

Equation (

12), and it uses the reset gate r

to monitor the inﬂow of the previous hidden

state h

t−1

, which contains past details. If the

reset value is set to 0, the previous state will

be restored. Hence, the reset gate provides a

method to to delete past secret states which are

not related to the future; that is, the reset gate

decides how much information was forgotten.

c) hidden state h

= z

∗ h

t−1

(

1 − z

)

∗

(13)

27850 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

FIGURE 6. LSTM layers Steps.

Hidden state (h

) is shown in Equation (

13).

It updates previous hidden state and the can-

didate hidden state ht − 1 with an update gate.

If the update gate z

is 1, the long-held previous

secret state h

can be moved to the current

moment. The GRU can handle the probability

gradient vanishing problem in the RNN, so it

is more suitable for detecting faults in dynamic

systems.

6) Evaluating models Four statistics are used to deter-

mine the consistency of models. TP is True Positive,

TN is True Negative, FP is False Positive, and FN is

False Negative. Accuracy is shown in Equation (

14)

and Precision is shown in Equation (

15). Recall is

shown in Equation (

16), and F1-Score is shown in

Equation (

17)

• Accuracy is a measure of totally correctly identi-

ﬁed samples out of all the samples.

Accuracy =

TP + TN

TP + TN + FP + FN

× 100 (14)

• Precision and Recall The measure of the ability of

the model to accurately identiﬁed the occurrence

of a positive class instance is determined by recall

Precision =

TP + FP

(15)

TABLE 6. The best values hyperparameter for the COVID 19 dataset for

Modified LSTM and Modified GRU.

Recall =

TP + FN

(16)

• F1-Score The harmonic mean of Precision and

Recall

F1 − Score =

2 ∗ Precision ∗ Recall

Precision + Recall

(17)

IV. EXPERIMENTAL RESULTS AND DISCUSSION

In this section, the results of applying six machine learn-

ing models (DT, LR, KNN, RF, SVM, NB) and Two deep

learning models (Modiﬁed LSTM, Modiﬁed GRU), includ-

ing cross-validation results and testing results, are described.

Each machine learning model performance is discussed using

four sizes of TF-IDF feature extraction, including uni-gram,

bi-gram, tri-gram, four-gram, and one matrix size; 3000.

In this section, the experimental evaluation of the proposed

models is presented, starting by describing the experiment

setup.

VOLUME 9, 2021 27851

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 7. The performance of ML for cross-validation results (COVID 19 Dataset).

A. EXPERIMENTAL SETUP

The ML and DL models are trained with 90% of the dataset

and then are tested with the 10% testing data. The six ML

classiﬁers are implemented using the sci-kit-learn package in

Python 3. The DL models are implemented using TensorFlow

and Keras package in Python 3.

B. EXPERIMENT I (COAID (COVID-19 HEALTHCARE

MISINFORMATION DATASET))

1) HYPERPARAMETER TUNING

The method of selecting hyperparameters is a vital part of

any deep learning solution. Deep learning algorithms all have

speciﬁcally deﬁned parameters that govern various aspects

of them. Hyperparameters are the variables that are set prior

to using a particular algorithm on a speciﬁc set of data.

The best numbers depend on each task and each context-

dependent dataset. The best values of parameters for the two

deep learning models (Modiﬁed LSTM, Modiﬁed GRU) are

shown in Table [

6].

2) CROSS-VALIDATION RESULT FOR THE

COVID 19 DATASET

We evaluate the performance of six machine learning mod-

els over a collection of datasets and show that, on aver-

age, the models predict 10 times better than random

guessing (3000).

TABLE 8. The performance of cross-validation for deep neural networks

(COVID 19 Dataset).

• the results of cross-validation are shown in Table [7]

with DT, LR, KNN, RF, SVM, NB, respectively using

3000 matrix size have obtained higher output for all

TF-IDF feature extraction methods, including uni-gram,

bi-gram, tri-gram and four-gram for machine learning

models and word embedding feature extraction method

for deep learning models. The Uni-gram model using

known text obtained the highest efﬁciency (accuracy

of 96.22%, the precision of 95.19%, recall of 96.23%

and F1-score 95.35%). Similar to KNN, the use of four-

grams resulted in the best results (accuracy of 95.87%,

the precision of 94.35%, recall of 95.87% and F1-score

94.66%). The RF model using bi-gram yielded the high-

est efﬁciency (accuracy of 96.63%, precision of 96.41%,

recall of 96.63% and F1-score 96.52%). As to LR,

the highest performances have been obtained using

bi (twice) gram (accuracy of 96.38%, the precision

of 95.73%, recall of 96.38% and F1-score 95.17%).

27852 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 9. The performance of ML for testing result (COVID 19 Dataset).

As the results show that the logistic regression model

with two coefﬁcients obtained the best efﬁciency

(accuracy of 96.64%, the precision of 96.45%, recall

of 96.64% and F1-score 95.53%). The NB model using

four-gram resulting in the highest efﬁciency (accuracy

of 96.24%, the precision of 95.67%, recall of 96.24%

and F1-score 94.73%).

• Regarding deep learning models, as shown in Table [

8]:

Modiﬁed LSTM, two layers, which have the highest

efﬁciency (accuracy of 98.57%, the precision of 98.82%,

recall of 99.71%, and F1-score 99.26%). Modiﬁed

GRU one layer, achieved the best efﬁciency (accuracy

of 98.33%, the precision of 98.67%, recall of 99.62%,

and F1-score 99.14%).

3) TESTING RESULTS FOR THE COVID 19 DATASET

In this section, we discuss the generalization performance of

the six machine learning models using the unseen test dataset

with matrix sizes (3000).

• As the results shown in Table [

9] described the testing

performance of machine learning models including DT,

LR, KNN, RF, SVM, NB, respectively.

. the uni-gram model using the DT technique obtained

the highest efﬁciency (accuracy of 95.0%, the preci-

sion of 93.71%, recall of 95.0% and F1-score 94.26%).

The KNN model using uni-gram obtained the highest

accuracy (accuracy of 95.66%, the precision of 92.16%,

recall of 95.66% and F1-score 93.81%). The RF bi-gram

model was the most effective; it had the greatest efﬁ-

ciency (accuracy of 96.35%, the precision of 96.07%,

recall of 96.35% and F1-score 95.0%). The LR model

that used four grams of text achieved the best overall

TABLE 10. The performance of testing results of deep neural networks

(COVID 19 Dataset).

TABLE 11. The best values hyperparameter of the Disasters Dataset for

Modified LSTM and Modified GRU.

efﬁciency (accuracy of 95.82%, precision of 94.24%,

recall of 95.82% and F1-score 94.59%). An unigram

SVM achieved the best overall efﬁciency (accuracy

of 96.36%, the precision of 96.18%, recall of 96.38%

and F1-score 95.05%). Using the tri-gram NB model,

obtained the highest score (accuracy of 96.13%, the pre-

cision of 95.74%, recall of 96.13% and F1-score 94.5%).

• Regarding deep learning models, as shown in

Table [

10]: Modiﬁed LSTM, two layers got the highest

results (accuracy of 98.6%, the precision of 98.55%,

recall of 98.6% and F1-score 98.5%). ‘Modiﬁed GRU,

a one layer system, achieved the best efﬁciency

(accuracy of 98.29%, the precision of 98.24%, recall

of 98.29% and F1-score 98.21%).

VOLUME 9, 2021 27853

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 12. The performance of ML for cross-validation results (Disasters Dataset).

TABLE 13. The performance of results deep neural networks for

cross-validation for the Disasters Dataset.

C. EXPERIMENT II: FAKE NEWS FOR (DISASTERS DATASET)

[45]

1) HYPERPARAMETER TUNING

The best values parameters for two deep learning models

(Modiﬁed LSTM, Modiﬁed GRU) are shown in Table [

11].

2) CROSS-VALIDATION RESULTS FOR DISASTERS DATASET

We experimentally demonstrate the performance of the

10-fold CV results of the six machine learning models over

the used Disasters Dataset with one matrix sizes (3000).

• As the results of cross-validation on the Disas-

ters Dataset shown in Table [

12] using DT, LR,

KNN, RF, SVM, NB, respectively, the dataset using

3000 matrix size have obtained higher performance

for all TF-IDF feature extraction methods, including

uni-gram, bi-gram, tri-gram and four-gram for machine

learning models and word embedding feature extrac-

tion method for deep learning models. We attribute

this behavior to the larger number of words within the

matrix. In particular, when the number of words is a

bit larger, the weighting metric becomes more signiﬁ-

cant, and this improves the machine learning and deep

learning models’ performance. However, the machine

learning models’ performance using 3000 matrix sizes

are varied based on the model and the feature extrac-

tion method. For example, Table [

12], the DT model

using uni-gram achieved the best efﬁciency (accuracy

of 75.3%, the precision of 75.58%, recall of 75.31,

and F1-score 74.83%). Similarly to the KNN model,

the best efﬁciency has been obtained using uni-gram

(accuracy of 77.03%, the precision of 78.44%, recall

of 77.03% and F1-score 76.08%). Similar to the RF

model, the best efﬁciency has been obtained using uni-

gram achieved the best efﬁciency (accuracy of 78.64%,

the precision of 79.41%, recall of 78.64% and F1-score

78.08%). Regarding LR, the highest performances have

been obtained using bi-gram achieved the best efﬁciency

(accuracy of 79.91%, the precision of 80.28%, recall

of 79.91% and F1-score 79.53). The SVM model has

obtained the best performance using Bi-gram (accuracy

of 80.08%, the precision of 80.69%, recall of 80.08,

and F1-score 79.61%). The NB has recorded the high-

est improvements using uni-gram and 3000 matrix size

27854 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 14. The performance of ML for testing results (Disasters Dataset).

(accuracy of 79.65%, the precision of 79.96%, recall

of 79.65 and F1-score 79.28%).

• • Regarding deep learning models, show in Table [13],

the dataset using the word embedding feature extraction

method for deep learning models obtained higher perfor-

mance for all layers, including one layer, two layers, and

three layers. The Modiﬁed LSTM one layer achieved

the best efﬁciency (accuracy of 82.68%, the preci-

sion of 85.86%, recall of 72.13and F1-score 78.14%).

Modiﬁed GRU one layer achieved the best efﬁciency

(accuracy of 80.37%, the precision of 83.37%, recall

of 70.93%, and F1-score 76.21%).

3) TESTING RESULTS FOR DISASTERS DATASET

• As the results of testing on the Disasters Dataset shown

in Table [

14] using DT, LR, KNN, RF, SVM, NB,

respectively. The DT model using uni-gram achieved

the best efﬁciency (accuracy of 68.42%, the precision

of 68.38%, recall of 68.42%, and F1-score 67.44%).

Similar to the KNN model, the best efﬁciency has

been obtained using uni-gram (accuracy of 66.68%,

the precision of 73.85%, recall of 66.68 and F1-score

61.16%). Similar to the RF model, the best efﬁciency

has been obtained using uni-gram (accuracy of 71.48%,

the precision of 75.08%, recall of 71.48, and F1-score

69.04%). Regarding LR, the highest performances have

been obtained using Four-gram (accuracy of 73.7%,

the precision of 76.65%, recall of 73.7 and F1-score

71.77%). The SVM model has obtained the best per-

formance using bi-gram obtained (accuracy of 72.9%,

the precision of 76.65%, recall of 72.9 and F1-score

70.6%). The NB has recorded the highest improve-

ments using Four-gram achieved the best efﬁciency

TABLE 15. The performance of testing results of deep neural networks on

the Disasters Dataset.

TABLE 16. The best values hyperparameter of the PolitiFact dataset for

Modified LSTM and Modified GRU.

(accuracy of 76.2%, the precision of 76.26%, recall

of 76.2 and F1-score 75.81%).

• • Regarding deep learning models, show in Table [

15],

the dataset using the word embedding Regarding deep

learning models. The Modiﬁed LSTM, one layer

achieved the best efﬁciency (accuracy of 86.74%,

the precision of 86.98%, recall of 86.74% and F1-score

86.6%). The Modiﬁed GRU one layer achieved the best

efﬁciency (accuracy of 81.44%, the precision of 81.68%,

recall of 81.44 % and F1-score 81.81%).

D. CASE III: (THE POLITIFACT DATASET) FACT-CHECKING

THE U.S POLITICAL NEWS

[46]

1) HYPERPARAMETER TUNING

The best values parameters for two deep learning models

(Modiﬁed LSTM, Modiﬁed GRU) are shown in Table [

16].

VOLUME 9, 2021 27855

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 17. The performance of ML for cross-validation results (The PolitiFact dataset).

TABLE 18. The performance of cross-validation for deep neural networks

(The PolitiFact dataset).

2) CROSS-VALIDATION RESULT FOR THE

POLITIFACT DATASET

We experimentally demonstrate the performance of the

10-fold CV results of the six machine learning models over

the used dataset with one matrix sizes (3000).

• As the results of cross-validation on the PolitiFact

dataset shown in Table [

17] using DT, LR, KNN,

RF, SVM, NB using 3000 matrix size have obtained

higher performance for all TF-IDF feature extrac-

tion methods, including uni-gram, bi-gram, tri-gram

and four-gram for machine learning models and word

embedding feature extraction method for deep learn-

ing models. The DT model using four-gram achieved

the best efﬁciency (accuracy of 76.8%, the precision

of 76.41%, recall of 76.8% and F1-score 76.73%).

Similar to the KNN model, using uni-gram achieved

the best efﬁciency (accuracy of 76.67%, the precision

of 79.96%, recall of 78.72% and F1-score 75.02%).

Similar to the RF model using bi-gram achieved the best

efﬁciency (accuracy of 78.15%, precision of 79.42%,

recall of 79.02% and F1-score 77.74%). Regarding LR,

the highest performances have been obtained using tri-

gram (accuracy of 81.91%, the precision of 81.94%,

recall of 81.91and F1-score 81.55%). The SVM model

using bi-gram achieved the best efﬁciency (accuracy

of 81.91%, the precision of 82.23%, recall of 82.02%

and F1-score 81.69%). The NB model using uni-

gram achieved the best efﬁciency (accuracy of 82.92%,

the precision of 83.01%, recall of 82.92% and F1-score

82.82%).

• Regarding deep learning models, as shown in Table [

18]:

The Modiﬁed LSTM, two layers achieved the best efﬁ-

ciency (accuracy of 94.21%, the precision of 96.15%,

recall of 88.0%, and F1-score 91.76%). The Modiﬁed

GRU one layer achieved the best efﬁciency (accuracy

of 88.8%, the precision of 87.85%, recall of 78.17%, and

F1-score 81.85%).

3) TESTING RESULTS FOR THE POLITIFACT DATASET

In this section, we discuss the generalization performance of

the six machine learning models using the unseen test dataset

with matrix sizes (3000).

• As the results shown in Table [

19] described the testing

performance of machine learning models including DT,

27856 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 19. The performance of ML for testing result (The PolitiFact dataset).

LR, KNN, RF, SVM, NB. The DT model using four-

gram achieved the best efﬁciency (accuracy of 75.08%,

the precision of 76.63%, recall of 75.17% and F1-score

74.16%). The KNN model using uni-gram achieved

the best efﬁciency (accuracy of 69.22%, the precision

of 70.88%, recall of 71.06% and F1-score 67.59%).

the RF model using bi-gram achieved the best efﬁ-

ciency (accuracy of 81.05%, the precision of 81.8%,

recall of 81.05% and F1-score 80.71%). the LR model

using tri-gram achieved the best efﬁciency (accuracy

of 79.0%, precision of 79.29%, recall of 79.4 and

F1-score 79.0%). the SVM model using Bi-gram

achieved the best efﬁciency (accuracy of 79.06%,

the precision of 79.8%, recall of 79.06% and F1-score

78.65%). The NB model using uni-gram achieved the

best efﬁciency (accuracy of 80.42%, the precision

of 80.68%, recall of 80.42% and F1-score 80.46%).

• Regarding deep learning models, as shown in [

20]:

The Modiﬁed LSTM, two layers achieved the best efﬁ-

ciency (accuracy of 83.93%, the precision of 86.66%,

recall of 83.93% and F1-score 83.31%). The Modiﬁed

GRU one layer achieved the best efﬁciency (accuracy

of 81.83%, the precision of 84.53%, recall of 81.83%

and F1-score 81.11%).

4) EXPERIMENT IV (GOSSIP COP DATASET (CHECKING THE

HOLLYWOOD AND CELEBRITY NEWS))

[46]

5) HYPERPARAMETER TUNING

The best values parameters for two deep learning models

(Modiﬁed LSTM, Modiﬁed GRU) are shown in Table [

21].

TABLE 20. The performance of testing results of deep neural networks

(The PolitiFact dataset).

TABLE 21. The best values hyperparameter of Gossip Cop Dataset for

Modified LSTM and Modified GRU.

6) CROSS-VALIDATION RESULTS FOR GOSSIP COP DATASET

In this section, the results of applying six machine learn-

ing models (D.T., L.R., KNN, RF, SVM, NB) and Two

deep learning models (The Modiﬁed LSTM, The Modi-

ﬁed GRU), including cross-validation results are described.

Each machine learning model performance is discussed using

four sizes of TF-IDF feature extraction, including uni-gram,

bi-gram, tri-gram, four-gram, and one matrix size; 3000.

In this section, the experimental evaluation of the proposed

models is presented, starting by describing the experiment

setup. We experimentally demonstrate the performance of the

10-fold CV results of the six machine learning models over

the Gossip Cop Dataset.

VOLUME 9, 2021 27857

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 22. The performance of ML for cross-validation result (Gossip Cop Dataset).

TABLE 23. The performance of cross-validation for deep neural networks (Gossip Cop Dataset).

• As the results shown in Table [22] described the Cross-

Validation performance of machine learning models

including DT, LR, KNN, RF, SVM, NB, respectively

on the third dataset, the DT model using four-gram

achieved the best efﬁciency (accuracy of 74.94%,

the precision of 75.06%, recall of 74.96% and F1-score

74.76%). The KNN model using uni-gram achieved

the best efﬁciency (accuracy of 77.62%, the precision

of 77.82%, recall of 77.62% and F1-score 77.37%).

The RF model using four-gram achieved the best efﬁ-

ciency (accuracy of 77.58%, the precision of 77.66%,

recall of 77.57% and F1-score 77.41%). The LR

model using four-gram achieved the best efﬁciency

(accuracy of 79.17%, the precision of 79.32%, recall

of 79.17% and F1-score 79.01%). The SVM model

using uni-gram achieved the best efﬁciency (accuracy

of 79.65%, the precision of 79.9%, recall of 79.65% and

F1-score 79.42%). The NB model using uni-gram

achieved the best efﬁciency (accuracy of 79.1%, the pre-

cision of 79.17%, recall of 79.1% and F1-score 79.03%).

• Regarding deep learning models, as shown in Table [

23]:

The Modiﬁed LSTM, one layer achieved the best efﬁ-

ciency (accuracy of 82.5%, the precision of 81.52%,

recall of 79.37% and F1-score 80.0%). The Modiﬁed

GRU two-layer achieved the best efﬁciency (accuracy

of 86.05%, the precision of 82.97%, recall of 75.66%

and F1-score 78.32%).

E. TESTING RESULT FOR THE GOSSIP COP DATASET

In this section, we discuss the generalization performance of

the six machine learning models using the unseen test dataset

with matrix sizes (3000).

• As the results show in Table [

24] described the testing

performance of machine learning models including DT,

27858 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

TABLE 24. The performance of ML for testing result (Gossip Cop Dataset).

LR, KNN, RF, SVM, NB, respectively, the DT model

using Uni-gram achieved the best efﬁciency (accuracy

of 71.74%, the precision of 72.05%, recall of 71.74%

and F1-score 71.74%). The KNN model using Tri-

gram achieved the best efﬁciency (accuracy of 71.2%,

the precision of 73.1%, recall of 71.2% and F1-score

of 70.6%). The RF model using Uni-gram achieved

the best efﬁciency (accuracy of 76.2%, the precision

of 76.91%, recall of 76.2% and F1-score 76.04%). The

LR model using Uni-gram achieved the best efﬁciency

(accuracy of 76.2%, precision 76.63%, recall of 76.2%,

and F1-score 76.1%). The SVM model using uni-gram

achieved the best efﬁciency (accuracy of 77.51%, preci-

sion 78.11%, recall of 77.51%, and F1-score 77.39%).

The NB model using uni-gram achieved the best efﬁ-

ciency (accuracy of 76.2%, precision 76.63%, recall

of 76.2%, and F1-score 76.1%).

• Regarding deep learning models, as shown in Table [

25]:

The Modiﬁed LSTM, one layer achieved the best efﬁ-

ciency (accuracy of 83.82%, precision 84.85%, recall

of 83.82%, and F1-score of 83.7%). The Modiﬁed

GRU two layers achieved the best efﬁciency (accu-

racy of 81.49%, precision 82.26%, recall of 81.49, and

F1-score 81.35%).

TABLE 25. The performance of ML for testing results (Gossip Cop

Dataset).

V. DISCUSSION

A. DISCUSSION FOR EXPERIMENT I(COVID-19 DATASET)

From the results obtained in our experiments for the

COVID-19 dataset, Figure

7 and Figure 8 depict the

empirical results in the big picture for the cross-validation

performances and the testing results, respectively. They are

showing the performance of the best models for each feature

extraction method. To summarize the performance of the

compared models, we explore the average cross-validation

and the testing results of each model using different sizes

of feature extraction methods, N-gram from n = 1 to

n = 4, and the matrix is 3000 for baseline machine learning

(DT, LR, KNN, RF, SVM, NB). The feature is extracted

using the word embedding method for Deep Neural Network

VOLUME 9, 2021 27859

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

FIGURE 7. The cross-validation performances results for the COVID-19 dataset.

FIGURE 8. The testing results for the COVID-19 dataset.

(The Modiﬁed LSTM, and The Modiﬁed GRU). On average,

The Modiﬁed LSTM model has obtained the best cross-

validation average and in the testing performance average

compared to other regular machine learning models. For

cross-validation results and testing results, the Modiﬁed

LSTM (Two layers) model has obtained the best an accu-

racy. for cross validation results; Accuracy 98.57%, precision

of 98.82%, recall of 99.71%, and F1-score of 99.26%. For

performance testing, The Modiﬁed LSTM has achieved an

accuracy of 98.6%, precision of 98.55%, recall of 98.6%,

and F1-score of 98.5%. The Modiﬁed GRU has obtained

the second-best rank of cross-validation and testing perfor-

mance using one layer. for cross validation results; (accuracy

of 98.33%, the precision of 98.67%, recall of 99.62%, and

F1-score of 99.14%). Also, it has been reported to be the

second-best results for testing (accuracy of 98.29%, the pre-

cision of 98.24%, recall of 98.29%, and F1-score of 98.21%).

SVM has obtained the third-best result on the average of

cross-validation performance, and in the testing results. For

cross-validation results, SVM using Bi-g ram has recorded

achieved (accuracy of 96.64%, precision 96.45%, recall

of 96.64%, and F1-score 95.53%). For testing results, SVM

has recorded using Uni-gram (accuracy of 96.38%, the preci-

sion of 96.18%, recall of 96.38% and F1-score 95.05%). RF

has obtained the fourth-best result on the average of cross-

validation performance and the testing results. For cross-

validation results, it has recorded using Bi-gram (accuracy

of 96.63%, the precision of 96.41%, recall of 96.63% and

F1-score 95.52%). For testing results, it has recorded using

Bi-gram (accuracy of 96.35%, the precision of 96.07%, recall

of 96.35% and F1-score 95.0%). NB has obtained the ﬁfth-

best result on the average of cross-validation performance,

and in the testing results. For cross-validation results, NB has

recorded achieved using Four-gram (accuracy of 96.24%,

the precision of 95.67%, recall of 96.24% and F1-score

94.73%). For testing results, using Tri-gram obtained (accu-

racy of 96.13%, the precision of 95.74%, recall of 96.13%,

and F1-score 94.5%). LR has obtained the sixth-best result on

the average of cross-validation performance and the testing

results. For cross-validation results, it has recorded using

27860 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

FIGURE 9. The cross-validation performances results for the disasters dataset.

FIGURE 10. The testing results for the disasters dataset.

Bi-gram (accuracy of 96.38%, the precision of 95.73%, recall

of 96.38% and F1-score 95.17%)). For testing results, it has

recorded using Four-gram (accuracy of 95.82%, the precision

of 94.24%, recall of 95.82%, and F1-score 94.59%). DT and

KNN have reported the lowest cross-validation and testing

results. DT achieved the seventh rank on the average of

cross-validation performance, and KNN achieved the best

in the testing results. For cross-validation results, DT has

recorded achieved using uni-gram (accuracy of 96.22%, pre-

cision 95.19%, recall of 96.23% and F1-score 95.35%). For

testing results, KNN has recorded using uni-gram achieved

the best efﬁciency (accuracy of 95.66%, precision 92.16%,

recall of 95.66% and F1-score 93.81%). KNN achieved the

last rank on the average of cross-validation performance,

and DT achieved the best in the testing results. For cross-

validation results, KNN has recorded achieved using Four-

gram (accuracy of 95.87%, the precision of 94.35%, recall

of 95.87% and F1-score 94.66%).; and for testing results,

DT has recorded using uni-gram (accuracy of 95.0%, the pre-

cision of 93.71%, recall of 95.0 % and F1-score 94.26%)

Consequently, The Modiﬁed LSTM and The Modiﬁed

GRU for the COVID-19 dataset are outperforming SVM, DT,

RF, NB, LR, and KNN for cross-validation and testing results.

Based on these results, it can be tentatively concluded that

The Modiﬁed LSTM and The Modiﬁed GRU classiﬁers will

be used in the fake news detection model.

B. DISCUSSION FOR EXPERIMENT II

(DISASTERS DATASET)

From the results obtained in our experiments for the disasters

dataset,Figure

9 and Figure 10 depict the empirical results in

the big picture for the cross-validation performances and the

testing results, respectively. On average, The Modiﬁed LSTM

(one layer) model has obtained the best cross-validation

average and the testing performance average compared to

other regular machine learning models. For cross-validation

results, the Modiﬁed LSTM model has achieved an accu-

racy of 82.68%, precision of 85.86%, recall of 72.13%, and

F1-score of 78.14% using one layer. For performance testing,

The Modiﬁed LSTM has achieved an accuracy of 86.74%,

VOLUME 9, 2021 27861

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

precision of 86.98%, recall of 86.74%, and F1-score of 86.6%

using also one layer. The Modiﬁed GRU has obtained the

second-best rank of cross-validation performance using one

layer (accuracy of 80.37%, the precision of 83.37%, recall

of 70.93%, and F1-score of 76.21%).

Also, it has been reported to be the second-best results

using two layers of performance testing (accuracy of 81.44%,

the precision of 81.68%, recall of 81.44%, and F1-score

of 81.81%).

SVM has obtained the third-best result on the average of

cross-validation performance, and NB achieved the best in

the testing results. For cross-validation results, SVM using

Bi-gram has recorded achieved (accuracy of 80.08%, preci-

sion 80.69%, recall of 80.08%, and F1-score 79.61%). For

testing results, NB has recorded using Four-gram (accuracy

of 76.2%, the precision of 76.26%, recall of 76.2% and

F1-score 75.81%).

LR has obtained the fourth-best result on the average

of cross-validation performance and the testing results. For

cross-validation results, it has recorded using Bi-gram (accu-

racy of 79.91%, the precision of 80.28%, recall of 79.91% and

F1-score 79.53%). For testing results, it has recorded using

Four-gram (accuracy of 73.7%, the precision of 76.65%,

recall of 73.7% and F1-score 71.77%).’ NB has obtained the

ﬁfth-best result on the average of cross-validation perfor-

mance, and SVM achieved the best in the testing results. For

cross-validation results, NB has recorded achieved using uni-

gram (accuracy of 79.65%, the precision of 79.96%, recall

of 79.65% and F1-score 79.28%). For testing results, SVM

has recorded using Bi-gram obtained (accuracy of 72.9%,

the precision of 76.65%, recall of 72.9%, and F1-score

70.6%).

RF has obtained the sixth-best result on the average of

cross-validation performance and the testing results. For

cross-validation results, it has recorded using uni-gram (accu-

racy of 78.64%, the precision of 79.41%, recall of 78.64% and

F1-score 78.08%)). For testing results, it has recorded using

uni-gram (accuracy of 71.48%, the precision of 75.08%,

recall of 71.48%, and F1-score 69.04%).

DT and KNN have reported the lowest cross-validation and

testing results.KNN achieved the seventh rank on the average

of cross-validation performance, and DT achieved the seventh

in the testing results. For cross-validation results, KNN has

recorded achieved using uni-gram (accuracy of 77.03%, pre-

cision 78.44%, recall of 77.03% and F1-score 76.08%). For

testing results, DT has recorded using uni-gram achieved the

best efﬁciency (accuracy of 68.42%, precision 68.38%, recall

of 68.42% and F1-score 67.44%). DT achieved the last rank

on the average of cross-validation performance, and KNN

achieved the best in the testing results. For cross-validation

results, DT has recorded achieved using uni-gram (accu-

racy of 75.3%, the precision of 75.58%, recall of 75.31%

and F1-score 74.83%).; and for testing results, KNN has

recorded using uni-gram (accuracy of 66.68%, the precision

of 73.85%, recall of 66.68 % and F1-score 61.16%).

Consequently, The Modiﬁed LSTM and The Modiﬁed

GRU for the disasters dataset are outperforming SVM, DT,

RF, NB, LR, and KNN for cross-validation and testing results.

Based on these results, it can be tentatively concluded that

The Modiﬁed LSTM and The Modiﬁed GRU classiﬁers will

be used in the fake news detection model.

FIGURE 11. The cross-validation performances results for the PolitiFact

dataset.

C. DISCUSSION FOR EXPERIMENT III

(POLITIFACT DATASET)

From the results obtained in our experiments for the PolitiFact

data set, Figure

11 and Figure 12 depict the empirical

results in the big picture for the cross-validation perfor-

mances and the testing results, respectively. On average, The

Modiﬁed LSTM (Two-layer) model has obtained the best

cross-validation average and the testing performance average

compared to other regular machine learning models.

For cross-validation results, the Modiﬁed LSTM model has

achieved an accuracy of 94.21%, precision of 96.15%, recall

of 88.0%, and F1-score 91.76%). For performance testing,

The Modiﬁed LSTM has achieved an accuracy of 83.93%,

precision of 86.66%, recall of 83.93%, and F1-score 83.31%).

The Modiﬁed GRU (One Layer) has obtained the second-

best rank of cross-validation performance using one layer

(accuracy of 88.8%, precision 87.85%, recall of 78.17%, and

F1-score 81.85%). Also, it has been reported to be the second-

best results of performance testing (accuracy of 81.83%,

precision 84.53%, recall of 81.83% and F1-score 81.11%).

NB has obtained the third-best result on the average of

cross-validation performance, and RF achieved the best in the

testing results.

For cross-validation results, NB using Uni-gram has

recorded achieved (accuracy of 82.92%, precision 83.01%,

recall of 82.92% and F1-score 82.82%). For testing results,

RF has recorded using Bi-gram (accuracy of 81.05%, pre-

cision 81.8%, recall of 81.05%, and F1-score 80.71%). LR

has obtained the fourth-best result on the average of cross-

validation performance, and NB achieved the best in the

testing results. For cross-validation results, LR using Bi-gram

27862 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

FIGURE 12. The testing results for the PolitiFact dataset.

has recorded achieved (accuracy of 81.91%, the precision

82.23%, recall of 82.02% and F1-score 81.69%). For test-

ing results, NB has recorded using the Uni-gram (accu-

racy of 80.42%, the precision of 80.68%, recall of 80.42%,

and F1-score 80.46%). LR has obtained the ﬁfth-best result

on the average of cross-validation performance, and SVM

achieved the best in the testing results. For cross-validation

results, LR has recorded achieved using Tri-gram (accuracy

of 81.91%, precision 81.94%, recall of 81.91% and F1-score

81.55%). For testing results, SVM has recorded using

Bi-gram obtained (accuracy of 79.06%, precision 79.8%,

recall of 79.06%, and F1-score 78.65%). RF has obtained

the sixth-best result on the average of cross-validation per-

formance, and LR achieved the best in the testing results.

For cross-validation results, RF has recorded achieved using

Bi-gram (accuracy of 78.15%, precision, 79.42%, recall

of 79.02% and F1-score 77.74%). For testing results, LR has

recorded using Tri-gram obtained (accuracy of 79.0%, preci-

sion 79.29%, recall of 79.4% and F1-score 79.0%).

DT and KNN have reported the lowest cross-validation and

testing results.KNN achieved the seventh rank on the average

of cross-validation performance, and DT achieved the best

in the testing results. For cross-validation results, KNN has

recorded achieved using uni-gram (accuracy of 76.67%, pre-

cision 79.96%, recall of 78.72% and F1-score 75.02%). For

testing results, DT has recorded using Four-gram achieved

the best efﬁciency (accuracy of 75.08%, precision 76.63%,

recall of 75.17% and F1-score 74.16%). DT achieved the

last rank on the average of cross-validation performance,

and KNN achieved the best in the testing results. For cross-

validation results, DT has recorded achieved using Four-

gram (accuracy of 76.8%, precision 76.41%, recall of 76.8%,

and F1-score 76.73%). For testing results, KNN has recorded

using uni-gram (accuracy of 69.22%, precision 70.88%,

recall of 71.06%, and F1-score 67.59%).

Consequently, The Modiﬁed LSTM and The Modiﬁed

GRU for the PolitiFact dataset are outperforming SVM, DT,

RF, NB, LR, and KNN for cross-validation and testing results.

Based on these results, it can be tentatively concluded that

The Modiﬁed LSTM and The Modiﬁed GRU classiﬁers will

be used in the fake news detection model.

FIGURE 13. The cross-validation performances results for the gossip cop

dataset.

FIGURE 14. The testing results for the gossip cop dataset.

D. DISCUSSION FOR CASE STUDY IV

(GOSSIP COP DATASET)

From the results obtained in our experiments for the gos-

sip cop dataset, ﬁgures

13 and 14 depict the empirical

results in the big picture for the cross-validation performances

and the testing results, respectively. On average, The Mod-

iﬁed GRU (Two-layer) model has obtained the best cross-

validation average, and The Modiﬁed LSTM (one layer) has

obtained the best testing performance average compared to

other regular machine learning models. For cross-validation

results, The Modiﬁed GRU (two layers) model has achieved

(accuracy of 86.05%, precision 82.97%, recall of 75.66%

and F1-score 78.32%). For performance testing, The Mod-

iﬁed LSTM (one layer) has achieved (accuracy of 83.82%,

precision 84.85%, recall of 83.82%, and F1-score 83.7%).

The Modiﬁed LSTM (One Layer) has obtained the second-

best rank of cross-validation performance(accuracy of 82.5%,

precision 81.52%, recall of 79.37%, and F1-score 80.0%).

Also, The Modiﬁed GRU (Two-layer) has been reported to

be the second-best testing results using (accuracy of 81.49%,

precision 82.26%, recall of 81.49%, and F1-score 81.35%).

SVM (Uni-Gram) has obtained the third-best result on the

average of cross-validation performance and testing results.

For cross-validation results, it recorded (accuracy of 79.65%,

VOLUME 9, 2021 27863

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

the precision of 79.9%, recall of 79.51%, and F1-score

79.42%). Furthermore, For testing results, it recorded (accu-

racy of 77.51%, the precision 78.11%, recall of 77.51%, and

F1-score 77.39%).

LR (Four Gram) has obtained the fourth-best result on

the average of cross-validation performance, and RF (Uni

Gram) achieved the best in the testing results. For cross-

validation results, LR using Four-gram has recorded achieved

(accuracy of 79.17%, precision 79.32%, recall of 79.17%

and F1-score 79.01%). For testing results, RF has recorded

using Uni-gram (accuracy of 76.2%, precision 76.91%, recall

of 76.2%, and F1-score 76.04%). NB has obtained the ﬁfth-

best result on the average of cross-validation performance,

and LR achieved the best in the testing results. For cross-

validation results, NB has recorded achieved using Uni-gram

(accuracy of 79.1%, precision 79.17%, recall of 79.1%, and

F1-score 79.03%). For testing results, LR has recorded using

Uni-gram obtained (accuracy of 76.2%, precision 76.63%,

recall of 76.2%, and F1-score 76.1 %). KNN has obtained

the sixth-best result on the average of cross-validation per-

formance, and NB achieved the best in the testing results. For

cross-validation results, KNN has recorded achieved using

Uni-gram (accuracy of 77.62%, precision of 77.82%, recall

of 77.62% and F1-score 77.37%). For testing results, NB has

recorded using Uni-gram obtained (accuracy of 76.2%, pre-

cision 76.63%, recall of 76.2%, and F1-score 76.1%). RF

achieved the Seventh rank on the average of cross-validation

performance, and DT achieved the best in the testing results.

For cross-validation results, RF has recorded achieved using

Four-gram (accuracy of 77.58%, precision of 77.66%, recall

of 77.57%, and F1-score 77.41%). For testing results, DT has

recorded using Uni-gram obtained (accuracy of 71.74%, pre-

cision 72.05%, recall of 71.74%, and F1-score 71.64%).

DT and KNN have reported the lowest cross-validation and

testing results respectively. DT (Four-Gram) achieved the

last rank on the average of cross-validation performance,

and KNN achieved the best in the testing results. For

cross-validation results, DT has recorded achieved using

four-gram (accuracy of 74.94%, precision 75.06%, recall

of 74.96% and F1-score 74.76%). For testing results, KNN

has recorded using Tri-gram achieved the best efﬁciency

(accuracy of 71.2%, precision 73.1%, recall of 71.2% and

F1-score 70.6%).

Consequently, The Modiﬁed LSTM and The Modiﬁed

GRU for the gossip cop dataset are outperforming SVM, DT,

RF, NB, LR, and KNN for cross-validation and testing results.

Based on these results, it can be tentatively concluded that

The Modiﬁed GRU and The Modiﬁed LSTM classiﬁers will

be used in the fake news detection model

VI. CONCLUSION AND FUTURE WORK

One of the most threatening events is the spread of COVID-19

virus. People seek trustworthy information from social media

to learn how to protect themselves. Misinformation can kill

people. In this paper, we proposed efﬁcient and enhanced

deep learning techniques to detect fake news from COVID-19

dataset and three other datasets (disasters, politifact and gos-

sip cop). All experiments are packed up by cross validation

and testing on unseen data to support the validity of our mod-

els. Regarding COVID-19 dataset, the best testing results are

obtained by LSTM (two layers). The performance measure

results are as follows: the accuracy is 98.6%, the precision is

98.55%, the recall is 98.6% and F1-score is 98.5%. Regard-

ing disasters dataset, the best testing results are obtained by

The Modiﬁed LSTM (one layer). The performance measure

results are as follows: the accuracy is 86.74%, the precision is

86.98%, the recall is 86.74% and F1-score is 86.6%. Regard-

ing politifact dataset, the best testing results are obtained by

The Modiﬁed LSTM (two layers). The performance mea-

sure results are as follows: the accuracy is 83.93%, the preci-

sion is 86.66%, the recall is 83.93% and F1-score is 83.31%.

Regarding gossip cop dataset, the best testing results are

obtained by The Modiﬁed LSTM (one layer). The perfor-

mance measure results are as follows: the accuracy is 83.82,

the precision is 84.85%, the recall is 83.82% and F1-score is

83.7%. We can conclude that the Modiﬁed LSTM with tuned

parameters (with one or two layers) proposed in this paper

outperformed DT, KNN, RF, LR, SVM, NB and baseline The

Modiﬁed LSTM and The Modiﬁed GRU models. The main

strength in our proposed approaches is the preprocessing

stage which depends on word embedding. Moreover, Some

parameters such as the number of neurons in each layer and

the drop out ratio greatly affect on the performance of the

deep learning technique. These parameters are optimized to

obtain the best performance. In this paper, the content features

are used in the binary classiﬁcation. In the future, we intend to

use a combination of content, temporal, and context features

to be used in multi-class classiﬁcation. Capsule networks

can be included in future plans to detect their effect on the

performance.

In this paper, we proposed optimized machine learning

and deep learning systems to detect fake news for COVID

19 and other data-sets. The prepossessing stage contained

an elaborate sentence analysis starting from removing unim-

portant characters till tokenization and stemming. Three dif-

ferent datasets are used and split into training and testing

sections. The feature analysis of machine learning approach

depends on TF-IDF and Ngrams, while the deep learning

approach depends on word embedding. Both approaches are

optimized using grid search and Keras tuning, respectively.

The performance of both approaches is measured using accu-

racy, precision, recall, and F1-measure. The deep learning

approaches outperformed the machine learning approaches in

the three datasets. However, The Modiﬁed LSTM (two lay-

ers) achieved the highest cross-validation accuracy (94.21%)

using the second dataset. While The Modiﬁed LSTM (one

layer) achieved the highest testing accuracy (86.74 %) using

the ﬁrst dataset. We recommend the Keras tuned The Mod-

iﬁed LSTM approach as a deep learning approach for fake

news detection.

27864 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

CONFLICTS OF INTEREST

The authors have declared that there is no conﬂict of interest.

Non-ﬁnancial competing interests.

AUTHOR CONTRIBUTIONS

All authors contributed equally to this paper, where Diaa

Salama participated in sorting the experiments, discussed

and analyzed the results, performed the experiments and

analyzed the results, wrote the paper, discussed the result,

and revised/edited the manuscript. Mohamed Taha, and

Ahmed Taha: performed the experiments and analyzed the

results s and wrote the paper. Fatma Helmy: discussed the

results and wrote the paper. Ayma Nabil: discussed the results

and revised the paper. Essam H Houssein: analyzed the results

and revised the paper. All authors reads and approved the

work in this paper.

REFERENCES

[1] L. Cui and D. Lee, ‘‘CoAID: COVID-19 healthcare misinfor-

mation dataset,’’ 2020, arXiv:2006.00885. [Online]. Available:

http://arxiv.org/abs/2006.00885

[2] CDC COVID-19 Response Team, ‘‘Severe outcomes among patients

with coronavirus disease 2019 (COVID-19)—United States, February 12-

March 16, 2020,’’ Morbidity Mortality Weekly Rep., vol. 69, no. 12,

pp. 343–346, 2020.

[3] N. K. Conroy, V. L. Rubin, and Y. Chen, ‘‘Automatic deception detection:

Methods for ﬁnding fake news,’’ Proc. Assoc. Inf. Sci. Technol., vol. 52,

no. 1, pp. 1–4, Jan. 2015.

[4] H. Allcott and M. Gentzkow, ‘‘Social media and fake news in the 2016 elec-

tion,’’ J. Econ. Perspect., vol. 31, no. 2, pp. 36–211, 2017.

[5] J. Z. Pan, S. Pavlova, C. Li, N. Li, Y. Li, and J. Liu, ‘‘Content based

fake news detection using knowledge graphs,’’ in Proc. Int. Semantic

Web Conf., Monterey, CA, USA. Cham, Switzerland: Springer, Oct. 2018,

pp. 669–683.

[6] V. L. Rubin, Y. Chen, and N. K. Conroy, ‘‘Deception detection for news:

Three types of fakes,’’ Proc. Assoc. Inf. Sci. Technol., vol. 52, no. 1, pp. 1–4,

Jan. 2015.

[7] K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, ‘‘Fake news detection

on social media: A data mining perspective,’’ ACM SIGKDD Explor.

Newslett., vol. 19, no. 1, pp. 22–36, 2017.

[8] W. Y. Wang, ‘‘‘Liar, liar pants on ﬁre’: A new benchmark dataset

for fake news detection,’’ 2017, arXiv:1705.00648. [Online]. Available:

http://arxiv.org/abs/1705.00648

[9] J. Thorne, A. Vlachos, C. Christodoulopoulos, and A. Mittal, ‘‘FEVER:

A large-scale dataset for fact extraction and VERiﬁcation,’’ 2018,

arXiv:1803.05355. [Online]. Available: http://arxiv.org/abs/1803.05355

[10] M. Potthast, S. Köpsel, B. Stein, and M. Hagen, ‘‘Clickbait detection,’’

in Proc. Eur. Conf. Inf. Retr., Padua, Italy. Cham, Switzerland: Springer,

Mar. 2016, pp. 810–817.

[11] G. C. Santia and J. R. Williams, ‘‘BuzzTace: A news veracity dataset with

Facebook user commentary and egos,’’ in Proc. 12th Int. AAAI Conf. Web

Social Media, 2018, pp. 531–540.

[12] E. Tacchini, G. Ballarin, M. L. D. Vedova, S. Moret, and

L. de Alfaro, ‘‘Some like it hoax: Automated fake news detection

in social networks,’’ 2017, arXiv:1704.07506. [Online]. Available:

http://arxiv.org/abs/1704.07506

[13] T. Mitra and E. Gilbert, ‘‘CREDBANK: A large-scale social media cor-

pus with associated credibility annotations,’’ in Proc. ICWSM, 2015,

pp. 258–267.

[14] K. Shu, D. Mahudeswaran, S. Wang, D. Lee, and H. Liu, ‘‘FakeNews-

Net: A data repository with news content, social context and spatial-

temporal information for studying fake news on social media,’’ 2018,

arXiv:1809.01286. [Online]. Available: http://arxiv.org/abs/1809.01286

[15] G. K. Shahi and D. Nandini, ‘‘FakeCovid—A multilingual cross-domain

fact check news dataset for COVID-19,’’ arXiv e-Prints, 2020.

[16] X. Zhou, A. Mulay, E. Ferrara, and R. Zafarani, ‘‘ReCOVery: A mul-

timodal repository for COVID-19 news credibility research,’’ 2020,

arXiv:2006.05557. [Online]. Available: http://arxiv.org/abs/2006.05557

[17] S. A. Memon and K. M. Carley, ‘‘Characterizing COVID-19 misinforma-

tion communities using a novel Twitter dataset,’’ 2020, arXiv:2008.00791.

[Online]. Available: http://arxiv.org/abs/2008.00791

[18] Y. Li, B. Jiang, K. Shu, and H. Liu, ‘‘MM-COVID: A multilingual and mul-

timodal data repository for combating COVID-19 disinformation,’’ 2020,

arXiv:2011.04088. [Online]. Available: http://arxiv.org/abs/2011.04088

[19] M. S. Al-Rakhami and A. M. Al-Amri, ‘‘Lies kill, facts save: Detect-

ing COVID-19 misinformation in Twitter,’’ IEEE Access, vol. 8,

pp. 155961–155970, 2020.

[20] R. Vijjali, P. Potluri, S. Kumar, and S. Teki, ‘‘Two stage transformer

model for COVID-19 fake news detection and fact checking,’’ 2020,

arXiv:2011.13253. [Online]. Available: http://arxiv.org/abs/2011.13253

[21] T. Hossain, R. L. Logan, IV, A. Ugarte, Y. Matsubara, S. Young, and

S. Singh, ‘‘COVIDLies: Detecting COVID-19 misinformation on social

media,’’ in Proc. 1st Workshop NLP COVID-19 (Part) EMNLP, 2020.

[22] M. K. Elhadad, K. F. Li, and F. Gebali, ‘‘Detecting misleading information

on COVID-19,’’ IEEE Access, vol. 8, pp. 165201–165215, 2020.

[23] L. Rundo, R. Pirrone, S. Vitabile, E. Sala, and O. Gambino, ‘‘Recent

advances of HCI in decision-making tasks for optimized clinical work-

ﬂows and precision medicine,’’ J. Biomed. Informat., vol. 108, Aug. 2020,

Art. no. 103479.

[24] A. Ghenai and Y. Mejova, ‘‘Fake cures: User-centric modeling of health

misinformation in social media,’’ Proc. ACM Hum.-Comput. Interact.,

vol. 2, pp. 1–20, Sep. 2018.

[25] J. S. Brennen, F. Simon, P. N. Howard, and R. K. Nielsen, ‘‘Types, sources,

and claims of COVID-19 misinformation,’’ Reuters Inst., vol. 7, pp. 1–3,

Apr. 2020.

[26] D. Kar, M. Bhardwaj, S. Samanta, and A. P. Azad, ‘‘No rumours please!

A multi-indic-lingual approach for COVID fake-tweet detection,’’ 2020,

arXiv:2010.06906. [Online]. Available: http://arxiv.org/abs/2010.06906

[27] H. Zhang, Z. Fan, J. Zheng, and Q. Liu, ‘‘An improving deception detection

method in computer-mediated communication,’’ J. Netw., vol. 7, no. 11,

p. 1811, Nov. 2012.

[28] S. Afroz, M. Brennan, and R. Greenstadt, ‘‘Detecting hoaxes, frauds, and

deception in writing style online,’’ in Proc. IEEE Symp. Secur. Privacy,

May 2012, pp. 461–475.

[29] E. J. Briscoe, D. S. Appling, and H. Hayes, ‘‘Cues to deception in

social media communications,’’ in Proc. 47th Hawaii Int. Conf. Syst. Sci.,

Jan. 2014, pp. 1435–1443.

[30] V. Pérez-Rosas and R. Mihalcea, ‘‘Experiments in open domain deception

detection,’’ in Proc. Conf. Empirical Methods Natural Lang. Process.,

2015, pp. 1120–1125.

[31] V. Rubin, N. Conroy, Y. Chen, and S. Cornwell, ‘‘Fake news or

truth? Using satirical cues to detect potentially misleading news,’’ in

Proc. 2nd Workshop Comput. Approaches Deception Detection, 2016,

pp. 7–17.

[32] B. D. Horne and S. Adali, ‘‘This just in: Fake news packs a lot

in title, uses simpler, repetitive content in text body, more similar to

satire than real news,’’ 2017, arXiv:1703.09398. [Online]. Available:

http://arxiv.org/abs/1703.09398

[33] C. Castillo, M. Mendoza, and B. Poblete, ‘‘Predicting information cred-

ibility in time-sensitive social media,’’ Internet Res., vol. 23, no. 5,

pp. 560–588, Oct. 2013.

[34] T. Bodnar, C. Tucker, K. Hopkinson, and S. G. Bilén, ‘‘Increasing the

veracity of event detection on social media networks through user trust

modeling,’’ in Proc. IEEE Int. Conf. Big Data (Big Data), Oct. 2014,

pp. 636–643.

[35] J. Ito, J. Song, H. Toda, Y. Koike, and S. Oyama, ‘‘Assessment of tweet

credibility with LDA features,’’ in Proc. 24th Int. Conf. World Wide Web,

May 2015, pp. 953–958.

[36] C. Castillo, M. Mendoza, and B. Poblete, ‘‘Information credibility on

Twitter,’’ in Proc. 20th Int. Conf. World Wide Web, 2011, pp. 675–684.

[37] M. Hardalov, I. Koychev, and P. Nakov, ‘‘In search of credible news,’’

in Proc. Int. Conf. Artif. Intell. Methodol., Syst., Appl., Varna, Bulgaria.

Cham, Switzerland: Springer, Sep. 2016, pp. 172–180.

[38] L. Zhou, D. P. Twitchell, T. Qin, J. K. Burgoon, and J. F. Nunamaker,

‘‘An exploratory study into deception detection in text-based computer-

mediated communication,’’ in Proc. 36th Annu. Hawaii Int. Conf. Syst. Sci.,

2003, p. 10.

[39] J. Ma, W. Gao, P. Mitra, S. Kwon, B. J. Jansen, K.-F. Wong, and M. Cha,

Detecting Rumors From Microblogs With Recurrent Neural Networks.

Palo Alto, CA, USA: AAAI Press, 2016.

VOLUME 9, 2021 27865

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

[40] J. Schmidhuber and S. Hochreiter, ‘‘Long short-term memory,’’ Neural

Comput, vol. 9, no. 8, pp. 1735–1780, 1997.

[41] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares,

H. Schwenk, and Y. Bengio, ‘‘Learning phrase representations using

RNN encoder-decoder for statistical machine translation,’’ 2014,

arXiv:1406.1078. [Online]. Available: http://arxiv.org/abs/1406.1078

[42] N. Ruchansky, S. Seo, and Y. Liu, ‘‘CSI: A hybrid deep model for fake news

detection,’’ in Proc. ACM Conf. Inf. Knowl. Manage., 2017, pp. 797–806.

[43] S. Volkova, K. Shaffer, J. Y. Jang, and N. Hodas, ‘‘Separating facts from

ﬁction: Linguistic models to classify suspicious and trusted news posts on

Twitter,’’ in Proc. 55th Annu. Meeting Assoc. Comput. Linguistics (Short

Papers), vol. 2, 2017, pp. 647–653.

[44] O. Ajao, D. Bhowmik, and S. Zargari, ‘‘Fake news identiﬁcation on Twitter

with hybrid CNN and RNN models,’’ in Proc. 9th Int. Conf. Social Media

Soc., Jul. 2018, pp. 226–230.

[45] G. Sarin and P. Kumar, ‘‘ConvGRUText: A deep learning method for fake

text detection on online social media,’’ in Proc. PACIS, 2020, p. 60.

[46] B. Guo, Y. Ding, L. Yao, Y. Liang, and Z. Yu, ‘‘The future of false

information detection on social media: New perspectives and trends,’’ ACM

Comput. Surv., vol. 53, no. 4, pp. 1–36, 2020.

[47] J. R. Quinlan, ‘‘Induction of decision trees,’’ Mach. Learn., vol. 1, no. 1,

pp. 81–106, 1986.

[48] T. K. Ho, ‘‘Random decision forests,’’ in Proc. 3rd Int. Conf. Document

Anal. Recognit., vol. 1, 1995, pp. 278–282.

[49] L. Breiman, ‘‘Random forests,’’ Mach. Learn., vol. 45, no. 1, pp. 5–32,

2001.

[50] T. Cover and P. Hart, ‘‘Nearest neighbor pattern classiﬁcation,’’ IEEE

Trans. Inf. Theory, vol. IT-13, no. 1, pp. 21–27, Jan. 1967.

[51] A. Stupar, S. Michel, and R. Schenkel, ‘‘Rankreduce–processing k-nearest

neighbor queries on top of mapreduce,’’ Large-Scale Distrib. Syst. Inf.

Retr., vol. 15, Jan. 2010, pp. 13–18.

[52] C.-W. Hsu, C.-C. Chang, and C.-J. Lin. (2003). A Practical Guide

to Support Vector Classiﬁcation. [Online]. Available: http://www.csie.

ntu.edu.tw/~cjlin/papers/guide/guide.pdf

[53] A. J. Smola and B. Schölkopf, ‘‘A tutorial on support vector regression,’’

Statist. Comput., vol. 14, no. 3, pp. 199–222, 2004.

[54] F. E. Harrell, ‘‘Ordinal logistic regression,’’ in Regression Modeling Strate-

gies. Cham, Switzerland: Springer, 2015, pp. 311–325.

[55] M. Pal, ‘‘Random forest classiﬁer for remote sensing classiﬁcation,’’ Int.

J. Remote Sens., vol. 26, no. 1, pp. 217–222, Jan. 2005.

[56] F. Chollet, ‘‘Building autoencoders in Keras,’’ The Keras Blog,

Tech. Rep., May 2016.

[57] N. K. Manaswi, ‘‘Understanding and working with Keras,’’ in Deep Learn-

ing With Applications Using Python. Berkeley, CA, USA: Apress, 2018,

pp. 31–43.

[58] C. Cortes, M. Mohri, and A. Rostamizadeh, ‘‘L2 regularization

for learning kernels,’’ 2012, arXiv:1205.2653. [Online]. Available:

https://arxiv.org/ftp/arxiv/papers/1205/1205.2653.

[59] P. Baldi and P. J. Sadowski, ‘‘Understanding dropout,’’ in Proc. Adv. Neural

Inf. Process. Syst., 2013, pp. 2814–2822.

[60] N. K. Manaswi, ‘‘CNN in Keras,’’ in Deep Learning With Applications

Using Python. Berkeley, CA, USA: Apress, 2018, pp. 105–114.

[61] A. Gulli and S. Pal, Deep Learning With Keras. Birmingham, U.K.: Packt

Publishing, 2017.

[62] A. F. Agarap, ‘‘Deep learning using rectiﬁed linear units (ReLU),’’ 2018,

arXiv:1803.08375. [Online]. Available: http://arxiv.org/abs/1803.08375

[63] Z. Zhang, ‘‘Improved adam optimizer for deep neural networks,’’ in

Proc. IEEE/ACM 26th Int. Symp. Qual. Service (IWQoS), Jun. 2018,

pp. 1–2.

[64] A. Wanto, A. P. Windarto, D. Hartama, and I. Parlina, ‘‘Use of binary

sigmoid function and linear identity in artiﬁcial neural networks for

forecasting population density,’’ Int. J. Inf. Syst. Technol., vol. 1, no. 1,

pp. 43–54, 2017.

[65] S. Kombrink, T. Mikolov, M. Karaﬁát, and L. Burget, ‘‘Recurrent neural

network based language modeling in meeting recognition,’’ in Proc. 12th

Annu. Conf. Int. Speech Commun. Assoc., 2011, pp. 2877–2880.

[66] X. Li and X. Wu, ‘‘Constructing long short-term memory based deep recur-

rent neural networks for large vocabulary speech recognition,’’ in Proc.

IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Apr. 2015,

pp. 4520–4524.

[67] G.-W. You, S. Park, and D. Oh, ‘‘Diagnosis of electric vehicle batteries

using recurrent neural networks,’’ IEEE Trans. Ind. Electron., vol. 64, no. 6,

pp. 4885–4893, Jun. 2017.

[68] F. A. Gers, J. Schmidhuber, and F. Cummins, Learning to Forget: Contin-

ual Prediction With LSTM. Edison, NJ, USA: IET, 1999.

[69] F. A. Gers, N. N. Schraudolph, and J. Schmidhuber, ‘‘Learning pre-

cise timing with lstm recurrent networks,’’ J. Mach. Learn. Res., vol. 3,

pp. 115–143, Aug. 2002.

[70] A. Graves, A.-R. Mohamed, and G. Hinton, ‘‘Speech recognition with deep

recurrent neural networks,’’ in Proc. IEEE Int. Conf. Acoust., Speech Signal

Process., May 2013, pp. 6645–6649.

[71] M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio, ‘‘Light gated recur-

rent units for speech recognition,’’ IEEE Trans. Emerg. Topics Comput.

Intell., vol. 2, no. 2, pp. 92–102, Apr. 2018.

[72] T. Cooijmans, N. Ballas, C. Laurent, Ç. Gülçehre, and A. Courville,

‘‘Recurrent batch normalization,’’ 2016, arXiv:1603.09025. [Online].

Available: http://arxiv.org/abs/1603.09025

[73] D. P. Kingma and J. Ba, ‘‘Adam: A method for stochastic

optimization,’’ 2014, arXiv:1412.6980. [Online]. Available:

http://arxiv.org/abs/1412.6980

DIAA SALAMA ABDELMINAAM received the

Ph.D. degree in information system from the

Faculty of Computers and Information, Menuﬁa

University, Egypt, in 2015. He has been an Assis-

tance Professor with the Information Systems

Department, Faculty of Computers and Informa-

tion, Benha University, Egypt, since 2011. He has

worked on several research topics. He has con-

tributed more than 40 technical articles in the areas

of wireless networks, wireless network security,

information security and internet applications, cloud computing, mobile

cloud computing, the Internet of Things, and machine learning in interna-

tional journals, international conferences, local journals, and local confer-

ences. He majors in cryptography, network security, the IoT, big data, cloud

computing, and deep learning.

FATMA HELMY ISMAIL graduated from the Fac-

ulty of Engineering, Ain Shams University. She

received the master’s and Ph.D. degrees from the

Faculty of Engineering, Ain Shams University.

She worked in database and software industries for

eight years. She is currently an Associate Professor

of computer science with Misr International Uni-

versity. Her research interests include data mining,

machine learning, computational intelligence and

swarm optimization, and deep learning.

MOHAMED TAHA received the M.Sc. and Ph.D.

degrees in computer science from Ain Shams Uni-

versity, Egypt, in February 2009 and July 2015,

respectively. He is currently an Assistant Pro-

fessor with the Computer Science Department,

Faculty of Computers and Artiﬁcial intelligence,

Benha University, Egypt. He has contributed more

than 20 technical articles in international journals

and conferences. His research interests include

computer vision (object tracking-video surveil-

lance systems), digital forensics (image forgery detection–document forgery

detection–fake currency detection), image processing (OCR), computer net-

work (routing protocols—security), augmented reality, cloud computing, and

data mining (association rules mining-knowledge discovery).

27866 VOLUME 9, 2021

D. S. Abdelminaam et al.: CoAID-DEEP: An Optimized Intelligent Framework

AHMED TAHA received the M.Sc. and Ph.D.

degrees in computer science from Ain Shams Uni-

versity, Egypt, in February 2009 and July 2015,

respectively. He is currently working as an

Assistant Professor with the Computer Sci-

ence Department, Benha University, Egypt. His

research interests include computer vision and

image processing (human behavior analysis -

video surveillance systems), digital forensics

(image forgery detection–document forgery detec-

tion), security (encryption–steganography–cloud computing), and content-

based retrieval (Arabic text retrieval—video scenes classiﬁcation-video

scenes retrieval–trademark image retrieval—closed-caption technology).

ESSAM H. HOUSSEIN received the Ph.D. degree

in computer science ‘‘wireless networks based on

artiﬁcial intelligence’’ in 2012. He is currently

working as an Associate Professor with the Fac-

ulty of Computers and Information, Minia Uni-

versity, Egypt. He is also the Founder of the

Computing and Artiﬁcial Intelligence Research

Group (CAIRG) in Egypt. He has more than

80 scientiﬁc research articles published in presti-

gious international journals in the topics of opti-

mization, machine learning, image processing, the IoT, and its applications.

He also serves as a reviewer for more than 30 journals (Elsevier, Springer,

and IEEE). His research interests include wireless sensor networks, the IoT,

bioinformatics and biomedical, cloud computing, soft computing, image

processing, artiﬁcial intelligence, data mining, optimization, and meta-

heuristics techniques.

AYMAN NABIL received the B.S. degree (Hons.)

from the Faculty of Computers and Informa-

tion, Helwan University, Egypt, in 2002, and

the master’s and Ph.D. degrees from the Faculty

of Computer and Information, Helwan Univer-

sity. He worked with Misr International Univer-

sity, Egypt. He is specializing in uncertainty, and

knowledge discovery approaches. He has pub-

lished articles under artiﬁcial intelligence area in

international journals, international conferences,

local journals, and local conferences.

VOLUME 9, 2021 27867